How RepliSec works under the hood.

Where does my data go?

Nowhere, unless you configure a cloud model and use the AI-assisted workflow policy. Concretely:

• Your documents and answers are written to disk in the directory you mounted. They are never uploaded anywhere by RepliSec.
• Embeddings are always computed locally. RepliSec uses sentence-transformers with the all-MiniLM-L6-v2 model (384 dimensions) running in-process. There is no configuration in which your text is sent to a third party to be embedded.
• The only outbound call RepliSec makes is to the LLM you configured, and only in a mode that permits LLM use. If you point it at Ollama on your own machine, that call does not leave your machine either.
• No telemetry. Chroma, Next.js and LiteLLM all ship usage telemetry enabled by default; all three are disabled in the image.

Do I need a cloud API key?

No. With a local model, or under the approved-answers-only and no-AI policies, RepliSec needs no API key and sends your content nowhere. These setups need no ongoing external service connection. Semantic matching requires a one-time embedding-model download before an environment is isolated (see Air-gapped deployment). No-AI lexical mode requires no model or API key at all. If you choose a cloud model, you supply your own key and pay that provider directly.

What are the answer policies?

The homepage describes three ways an answer can be produced:

• AI-assisted workflow: AI may draft wording or provide reference material; a person reviews and approves every answer. In Human Authored mode, a person writes the final response and remains the author of record.
• Approved answers only: an existing human-approved answer is returned verbatim. Semantic matching may be used to find it, but nothing is generated.
• No AI: deterministic lexical matching. No LLM, embeddings or neural model is loaded.

The policy is set for the engagement, so every answer in a questionnaire follows the same rules.

How do they map to the compliance modes in the product?

Inside the product, and in its configuration, the three policies are implemented as five compliance modes:

Human Authored belongs to the AI-assisted workflow because a model may still provide reference material. The final response, though, is written by a person, who remains the author of record.

Is semantic matching "no AI"?

No. Semantic matching uses embeddings, which are produced by a small neural model (all-MiniLM-L6-v2) running locally. That is AI, used to find text by meaning. It is not generative: it ranks your existing, approved answers and never writes or changes wording. Generative drafting is the separate step where an LLM composes new text.

So the approved-answers-only policy is no generative drafting, not no AI. Only Library Only (Lexical) is no AI.

How is "no AI" enforced?

When RepliSec boots with REPLISEC_COMPLIANCE_MODE_DEFAULT=library_only_lexical, the modules that would import sentence-transformers and litellm are short-circuited at import time. Those libraries never enter the process. It is not that the code declines to call them. The code that could call them is not there.

Every other mode enforces its rules per call: each entry point that could reach a model asks the gate first.

How do I use it with a local LLM?

RepliSec talks to models through LiteLLM, so anything LiteLLM can reach, it can use. For a fully local setup, Ollama is the shortest path.

1. Install and start Ollama on your host (not in the container):

   brew install ollama          # macOS; see ollama.com for Linux/Windows
   ollama serve

2. Pull a model. An 8B model is a reasonable starting point on a laptop:

   ollama pull llama3.1:8b

3. Start RepliSec. Nothing special is needed, and no API key:

   docker run -d --rm --name replisec \
     -p 3000:3000 -v $(pwd)/data:/app/data \
     -e REPLISEC_ENCRYPTION_KEY=$(openssl rand -base64 32) \
     replisec:latest

4. Point RepliSec at Ollama. Open http://localhost:3000, go to Settings → AI Provider, and choose Ollama (local). Two fields:

   • Ollama URL: http://host.docker.internal:11434 on macOS and Windows. On Linux, use http://172.17.0.1:11434, or run the container with --network host and use http://localhost:11434.
   • Model: llama3.1:8b, or whatever you pulled; the name must match ollama list exactly.

   There is no API key field, because there is no API key.

5. Test it. Settings has a Test connection button. Then upload a document and generate answers on a questionnaire. The whole pipeline is now local: local embeddings, local vector store, local model.

Other local runtimes. LM Studio, vLLM, LocalAI and anything else exposing an OpenAI-shaped API work through the OpenAI-compatible provider. Set the base URL (for example http://localhost:1234/v1) and the model name the endpoint advertises.

Which models should I use?

For cloud providers, any current model from Anthropic, OpenAI, Google, Groq or Together works; you supply the model identifier and your key.

For local models, the trade-off is quality against hardware. An 8B model is usable for drafting from retrieved sources, which is the task here: the model is summarising your own material rather than recalling facts. Larger models produce better prose but need more memory.

Retrieval quality does not depend on the LLM at all. Finding the right source is done by the embedding model and the keyword index, both local and both fixed. The LLM only phrases the answer.

How are my provider credentials stored?

API keys are encrypted with AES-GCM before being written to SQLite, using a key you supply as REPLISEC_ENCRYPTION_KEY. That key is never persisted; it lives only in the container's environment.

If you rotate or lose it, RepliSec does not fail to start. It detects that the stored credentials can no longer be decrypted, treats them as absent, and asks you to re-enter them in Settings.

What is exposed on the network?

Only the UI on port 3000. The FastAPI backend is bound to loopback inside the container and reached through the UI's /api/* proxy. The container runs as a non-root user.